Akhilesh Keshap
I am a software engineer by education, but a problem solver by attitude. In my 12 years of experience spanning across various industries such as E-commerce, Cloud, SaaS Technology, Payments, Investment banking and Fintech, I have worked and managed large-scale, complex and multi-geographic projects in Technology and built products that customers love. During this time, I have worn multiple hats such as Product Manager, Go to Market strategy and execution, Team leader, Coach, etc.
Session
Modern AI agents are becoming increasingly capable, but reliability often gets worse as we give the LLM more tools, more decisions, and more responsibility.
This talk explores a different approach: designing agentic systems by deliberately reducing what the LLM has to decide.
We will look at three practical architecture patterns. First, how tool abstraction can reduce the decision surface by exposing meaningful capabilities instead of dozens of low-level APIs and MCP tools. Second, how to define reasoning boundaries—using deterministic code and rules where results can be computed, and reserving LLM reasoning for places where interpretation is genuinely required. Finally, we will look at how to evaluate these systems layer by layer instead of judging only the final response.
Along the way, we will cover structured evidence, typed contracts, failure boundaries, recommendation versus execution, and when LLM-as-a-judge is—and is not—the right evaluation technique.
Attendees will leave with a practical framework for deciding what the LLM should see, what it should decide, and how each part of an agentic workflow should be tested.